Testing a 3-Trillion Parameter MoE Model: Top-Tier Performance at One-Fifth the Cost
This article analyzes a newly released domestic Mixture-of-Experts (MoE) large language model featuring 3 trillion total parameters, a 1-million token context window, and multimodal capabilities. By activating only 104 billion parameters per inference, the model significantly reduces computational costs, allowing its API pricing to be set at one-fifth of top-tier closed-source alternatives. The review covers the model's technical architecture, benchmark performance in coding and agentic tasks, cost strategy, and the broader implications for the open-weight versus closed-source model landscape.
In recent years, the competition in the large language model (LLM) sector has focused on two main dimensions: the continuous scaling of parameters and intelligence, and the practical balance between inference costs and commercial deployment. Recently, a leading domestic AI provider released its latest Mixture-of-Experts (MoE) LLM. Combining 3 trillion total parameters, a 1-million token context window, and a price point set at just one-fifth of top-tier closed-source models, it has quickly captured the attention of the developer community. This article provides an in-depth analysis of the model across four aspects: technical architecture, core benchmarks, cost logic, and industry impact.
01 Architecture Analysis: The Cost Subtraction Behind 3 Trillion Parameters
First, it is essential to clarify the parameter metric. The officially claimed “3 trillion parameters” does not refer to a traditional dense model but rather the total parameter count of a Mixture-of-Experts (MoE) architecture. Based on公开 public technical details, the model has approximately 2.8 trillion total parameters, but only activates about 104 billion parameters during actual inference. This is the core value of the MoE architecture: maintaining a massive “knowledge skeleton” while only mobilizing the necessary “neurons” for each computation. This keeps computational overhead per inference extremely low while sustaining high intelligence levels.
This design directly underpins the model’s pricing strategy. For developers, understanding this distinction is critical: in API scenarios, costs are typically tied to the number of activated parameters (compute volume) rather than total parameters. A trillion-parameter MoE model can have a significantly lower single-inference cost than a 100-billion-parameter dense model. This structural advantage is why the model can price its API at one-fifth of leading closed-source alternatives. Additionally, the model natively supports visual understanding (multimodal input) and offers a 1-million token context window, sufficient for analyzing medium-sized code repositories or processing long-form documents.
Notably, prior to its official release, the model was tested anonymously on third-party evaluation platforms, undergoing public blind testing by global developers. Only after gaining initial recognition did the developers officially claim its identity on launch day. This “test first, announce later” strategy reflects a growing emphasis on real-world scenario evaluations, reversing the traditional path of releasing a technical report before launching the product.
02 Benchmark Deep Dive: Overtaking in Coding and Long-Horizon Tasks
Benchmark data is the true test of a model’s caliber. In several authoritative benchmarks published officially, the model demonstrated significant differentiated advantages, particularly in agent and coding scenarios. The table below excerpts key metric comparisons:
| Benchmark | New Model | Top-Tier Closed-Source Model A | Top-Tier Closed-Source Model B |
|---|---|---|---|
| Terminal-Bench 4.0 (Terminal Coding) | 65.7% | 58.2% | 66.4% |
| SWE-Marathon v1.1 (Long-Horizon SE) | 51.7% | 42.5% | 46.3% |
| Agents’ Last Exam (Agentic Comprehensive) | 43.2 | 32.8 | 42.6 |
| BenchCAD (3D Scene Reconstruction) | 92.7% | 95.9% | 96.2% |
Highlight 1: Coding Agent Capabilities Approaching the Strongest Closed-Source Models. In Terminal-Bench 4.0, which evaluates command execution in real terminal environments, the model scored 65.7%, leading Top-Tier Closed-Source Model A by nearly 7.5 percentage points and trailing the widely recognized coding leader, Model B, by just 0.7 percentage points. This indicates that in real-world development scenarios such as automated ops, script generation, and CLI tool invocation, the model possesses top-tier tool utilization and task planning capabilities.
Highlight 2: Significant Advantages in Long-Horizon Tasks. In SWE-Marathon v1.1 (simulating cross-file, multi-step software engineering marathon tasks), the model scored 51.7%, surpassing both competing closed-source models. This demonstrates its ability not only to generate isolated code snippets but also to understand project-level context and maintain long-horizon consistency—a core metric for building automated coding agents.
Highlight 3: Cross-Verification by Independent Evaluation Institutions. The independent evaluation agency Artificial Analysis awarded the project an Intelligence Index score of 55, ranking fourth globally and achieving an 11-point leap from its predecessor (44 points). The agency specifically noted that once model weights are open, it will become the highest-scoring open-weight model on their list, drastically narrowing the comprehensive gap between open-weight and closed-source models from the current 12 points to just 3 points. Although shortcomings remain in scientific computing terminal tasks and certain vertical knowledge domains, its overall capability distribution is highly balanced.
03 Pricing Strategy: The Catfish Effect Driving “Intelligence Equality”
Pricing is another core element drawing widespread attention to this model. Compared to functionally comparable top-tier closed-source models, its API call price is only one-fifth. This strategy is not simple low-price dumping but precise cost control based on architectural advantages. For startups and small-to-medium enterprises, this means they can integrate powerful AI capabilities into their business pipelines at a much lower trial-and-error cost.
From an industry perspective, this move will accelerate the “commoditization” and “equalization” of LLM capabilities. When the intelligence of top-tier models is no longer strictly tied to exorbitant API fees, developers can refocus on business scenario innovation and product experience refinement. Simultaneously, this puts substantial pricing pressure on closed-source model providers, pushing the entire industry to find a new equilibrium between performance and cost.
04 Landscape Outlook: The Tipping Point in the Open-Weight vs. Closed-Source Race
The release of this model is not merely an iteration of a single product, but a microcosm of the evolving LLM industry landscape. Over the past year, the debate over whether open-source or open-weight models can catch up to closed-source models in performance has never ceased. Current trends indicate that, thanks to the popularization of the MoE architecture, refined training data governance, and advancements in inference optimization technologies, the performance ceiling of open-weight models is being rapidly pushed higher.
The strength demonstrated by this new model validates a feasible path: utilizing extreme engineering capabilities to push model performance into the global top tier under finite cost constraints. Although there remains a marginal final step to fully surpassing top-tier closed-source models, its leading performance in specific critical tasks (such as coding agents) and its overwhelming price advantage already provide developers with a highly attractive alternative.
For technical decision-makers, model selection is no longer a single-dimensional “performance takes all” game. The industry has entered a multi-dimensional decision-making era requiring comprehensive consideration of intelligence level, inference cost, deployment flexibility (e.g., support for local private deployment), and data security compliance. As open-weight models increasingly close the gap, there is reason to anticipate the accelerated arrival of a more diverse and dynamic AI application ecosystem.
